Power substation equipment health state evaluation method and system based on multi-source monitoring data fusion

Abstract

The application discloses a substation equipment health state evaluation method and system based on multi-source monitoring data fusion, and relates to the technical field of smart power grids.The method comprises the following steps: loading multi-source heterogeneous state data, combining a transformer heat conduction partial differential equation to obtain a physical constraint loss value and a physical prior embedding vector, performing space-time graph Transformer fusion according to a transformer physical topology structure to obtain a cross-modal fusion feature vector, performing multi-task health evaluation and Bayesian uncertainty quantification to obtain confidence intervals of health grade classification probability distribution, hotspot temperature prediction value and residual life prediction value, and generating an explainability evaluation report based on an attention weight matrix and a contribution proportion of each equation term in the physical constraint loss value.The method solves the technical problems of poor explainability, incomplete output information, lack of confidence boundary, lack of self-adaptive updating capability of the model and failure to consider the edge end lightweight deployment demand in the existing evaluation process.

Description

Technical Field

[0001] This application relates to the field of smart grid technology, specifically to a method and system for assessing the health status of substation equipment by fusing multi-source monitoring data. Background Technology

[0002] With the advancement of smart grid and ubiquitous power Internet of Things construction, the number of various monitoring sensors configured on power equipment in substations is constantly increasing, accumulating massive amounts of multi-source, heterogeneous equipment operation and status data. As the core equipment of a substation, the health status of the transformer directly affects the operational safety and stability of the entire power grid. Accurately assessing the health status of transformers and predicting potential faults in advance is of great significance for substation operation, maintenance, and repair.

[0003] Existing transformer health status assessment methods, such as CN121071754A, embed the heat conduction equation and gas decomposition kinetic equation into the loss function through a physical information neural network, achieving coupled modeling of the temperature field and gas concentration. However, the assessment process is a black box, failing to clearly define the specific contributions of each sensor's data and physical laws to the assessment conclusions, making it difficult for maintenance personnel to trust the model's judgment. Similarly, CN120317527A and CN120930508B employ knowledge graphs and multi-branch networks for fault diagnosis or degradation trajectory prediction, but their outputs are limited, only providing the fault type or degradation trend, lacking the ability to simultaneously assess multiple tasks such as health level, hotspot temperature, and remaining lifespan. Furthermore, they fail to quantify the uncertainty of the prediction results and cannot provide assessment confidence intervals.

[0004] In summary, in the practical application of substation equipment health management, existing methods have poor interpretability in the assessment process, failing to meet the power system's requirements for decision-making transparency; the output information is incomplete, making it difficult to support multi-dimensional operation and maintenance decisions; the lack of confidence boundaries increases the risk of misjudgment; the assessment model lacks adaptive update capabilities, failing to adapt to the state drift of equipment during long-term operation, and does not consider the need for lightweight deployment at the edge. Summary of the Invention

[0005] This application provides a method and system for assessing the health status of substation equipment by fusing multi-source monitoring data. This solves the technical problems of poor interpretability of existing assessment processes, incomplete output information, lack of confidence boundaries, lack of adaptive update capabilities of models, and failure to consider the lightweight deployment requirements of edge terminals.

[0006] The technical solution to the above-mentioned technical problems in this application is as follows:

[0007] Firstly, this application provides a method for assessing the health status of substation equipment by fusing multi-source monitoring data, the method comprising:

[0008] Load multi-source heterogeneous state data of the transformer;

[0009] Based on the multi-source heterogeneous state data and the transformer heat conduction partial differential equation, physical information neural network encoding is performed to obtain physical constraint loss value and physical prior embedding vector;

[0010] Based on the multi-source heterogeneous state data and the physical prior embedding vector, spatiotemporal graph Transformer fusion is performed according to the physical topology of the transformer to obtain cross-modal fusion feature vectors;

[0011] Based on the cross-modal fusion feature vector, a multi-task health assessment is performed to simultaneously obtain the health level classification probability distribution, hotspot temperature prediction value, and remaining life prediction value.

[0012] Based on the model parameters in the multi-task health assessment, Bayesian uncertainty quantification is performed to obtain the confidence intervals of the health level classification probability distribution, the hotspot temperature prediction value, and the remaining lifespan prediction value.

[0013] Based on the attention weight matrix in the spatiotemporal graph Transformer fusion and the contribution ratio of each equation term in the physical constraint loss value, physical attention attribution is performed to generate an interpretability assessment report.

[0014] Secondly, this application provides a substation equipment health status assessment system based on multi-source monitoring data fusion, including:

[0015] The initial data acquisition module is used to load multi-source heterogeneous state data of the transformer;

[0016] The neural network encoding module is used to perform physical information neural network encoding based on the multi-source heterogeneous state data and the transformer heat conduction partial differential equation to obtain physical constraint loss value and physical prior embedding vector.

[0017] The feature vector fusion module is used to perform spatiotemporal graph Transformer fusion based on the multi-source heterogeneous state data and the physical prior embedding vector, according to the transformer physical topology, to obtain cross-modal fused feature vectors.

[0018] The multi-task health assessment module is used to perform multi-task health assessment based on the cross-modal fusion feature vector, and simultaneously obtain the health level classification probability distribution, hotspot temperature prediction value and remaining life prediction value.

[0019] The Bayesian quantization execution module is used to perform Bayesian uncertainty quantification based on the model parameters in the multi-task health assessment, and to obtain the confidence intervals of the health level classification probability distribution, the hotspot temperature prediction value, and the remaining life prediction value, respectively.

[0020] The evaluation report generation module is used to perform physical attention attribution and generate an interpretable evaluation report based on the attention weight matrix in the spatiotemporal graph Transformer fusion and the contribution ratio of each equation term in the physical constraint loss value.

[0021] This application provides one or more technical solutions, which have at least the following technical effects or advantages:

[0022] This application provides a method and system for assessing the health status of substation equipment through multi-source monitoring data fusion. First, the physical laws of transformer heat conduction are embedded into the neural network encoding process in the form of physical constraint loss. A spatiotemporal graph Transformer is constructed based on the actual physical topology of the transformer to achieve cross-modal fusion of multi-source heterogeneous data. This utilizes prior physical knowledge to improve the model's generalization ability and ensures assessment accuracy through data-driven feature fusion. Second, a multi-task health assessment with shared parameters is adopted, simultaneously outputting health level classification probabilities, hotspot temperature predictions, and remaining life predictions, achieving multi-dimensional health status output. Third, a Bayesian neural network is used to model cognitive uncertainty and random uncertainty for model parameters and data noise, providing clear confidence intervals for each output result and reducing the risk of misjudgment. Finally, a natural language interpretable assessment report is generated through physical attention attribution, clarifying the contribution of each sensor data and physical equation term to the assessment results, meeting the power system's requirements for decision-making transparency. Simultaneously, incremental learning enables dynamic updates of model parameters, and knowledge distillation facilitates cloud-edge collaborative deployment, adapting to the real-time assessment needs of substation edge environments.

[0023] Through the above technical solutions, this application solves the problems of insufficient interpretability, single output dimension, lack of result confidence description, inability to dynamically update the model, and difficulty in adapting to edge deployment in existing assessment methods. It can provide transparent, comprehensive, and reliable health status assessment support for substation transformer operation and maintenance, effectively improve the intelligence level of substation equipment operation and maintenance, and ensure the stability and security of power grid operation. Attached Figure Description

[0024] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0025] Figure 1 This is a flowchart illustrating the substation equipment health status assessment method based on multi-source monitoring data fusion provided in this application embodiment;

[0026] Figure 2 This is a flowchart illustrating the process of obtaining physical constraint loss values ​​and physical prior embedding vectors in the substation equipment health status assessment method based on multi-source monitoring data fusion provided in this application embodiment.

[0027] Figure 3 This is a schematic diagram of the substation equipment health status assessment system based on multi-source monitoring data fusion provided in this application embodiment.

[0028] The components represented by each number in the attached diagram are explained below:

[0029] The module includes an initial data acquisition module 11, a neural network encoding module 12, a feature vector fusion module 13, a multi-task health assessment module 14, a Bayesian quantization execution module 15, and an assessment report generation module 16. Detailed Implementation

[0030] This application provides a method and system for assessing the health status of substation equipment by fusing multi-source monitoring data. This is intended to address the technical problems of poor interpretability of existing assessment processes, incomplete output information, lack of confidence boundaries, lack of adaptive update capabilities of models, and failure to consider the need for lightweight deployment at the edge.

[0031] Example 1, as Figure 1 As shown in the embodiments of this application, a method for assessing the health status of substation equipment based on multi-source monitoring data fusion is provided, including:

[0032] S10: Multi-source heterogeneous state data of the loaded transformer;

[0033] In this embodiment, transformer data with different data modes and sampling frequencies are first collected by multiple sensors from different monitoring terminals. The data are then aligned according to time steps and standardized to form multi-source heterogeneous state data.

[0034] Specifically, step S10 in the method includes:

[0035] Obtain online monitoring data, operational data, environmental data, and ledger data of the transformer;

[0036] The online monitoring data, the operational data, the environmental data, and the ledger data are denoised, missing value imputed, and standardized, and then merged into the multi-source heterogeneous state data.

[0037] In this embodiment, the transformer's online monitoring data, operational data, environmental data, and ledger data are first acquired. The online monitoring data is continuously collected in real time by sensors and includes vibration data, oil temperature data, load current data, voltage data, and dissolved gas analysis data. The operational data includes active power data, reactive power data, and tap changer position data, reflecting the equipment's operating status. The environmental data includes ambient temperature data and cooler operating status data, reflecting the equipment's external environmental conditions. The ledger data refers to static archive data throughout the equipment's entire lifecycle, including equipment parameter data and historical maintenance records.

[0038] Secondly, the online monitoring data, operational data, environmental data, and ledger data are aligned in the time dimension. The sliding window mean method is used to denoise outliers caused by abnormal fluctuations in the sensors. The adjacent time step interpolation method is used to fill in occasional data interruptions and missing data. Finally, the numerical range of data with different dimensions is unified by Z-score standardization to complete the integration of multi-source heterogeneous state data.

[0039] Specifically, the sliding window mean method involves setting a fixed-length time window, taking the average value of all monitored data points within the window to replace the original value of the window's center point, and sliding the window one by one to complete the entire sequence processing, thereby filtering out random noise interference; the adjacent time step interpolation method is used for data segments with consecutive missing lengths less than a set threshold, and linear interpolation is performed using the effective data points at both ends to obtain the imputation value; Z-score standardization maps data of different dimensions to similar numerical intervals by subtracting the overall mean of the corresponding modality data from each data point and then dividing by the overall standard deviation of the corresponding modality, thus eliminating the impact of dimensional differences on model training.

[0040] Based on the above step S10, unified and regular multi-source heterogeneous state data can be obtained, providing a high-quality input foundation for subsequent feature encoding and fusion, and avoiding the interference of noise, missing data and dimensional differences in the original data on the model training and evaluation results.

[0041] S20: Based on the multi-source heterogeneous state data and the transformer heat conduction partial differential equation, perform physical information neural network encoding to obtain physical constraint loss value and physical prior embedding vector;

[0042] In this embodiment of the application, based on the multi-source heterogeneous state data obtained by loading as described above, the transformer heat conduction partial differential equation is embedded into the neural network as a physical constraint, so that the network output not only fits the observed data, but also strictly follows the laws of thermodynamics.

[0043] Specifically, step S20 in the method includes:

[0044] The transformer heat conduction partial differential equation is used as the core of physical constraints, and initial condition constraints and boundary condition constraints are incorporated to construct a physical constraint loss function.

[0045] The physical parameters of the transformer are encoded into learnable embedding vectors, which are set as physical prior embedding vectors. The physical parameters include specific heat capacity, thermal conductivity and thermal diffusivity.

[0046] In the calculation of the physical constraint loss function, an attention mechanism based on residuals is introduced to dynamically adjust the contribution weight of different residual points to the total loss function, thereby obtaining the physical constraint loss value.

[0047] In the embodiments of this application, such as Figure 2 As shown, firstly, the transformer heat conduction partial differential equation is incorporated into the initial condition constraints and boundary condition constraints to construct a physical constraint loss function. Specifically, the transformer heat conduction partial differential equation describes the evolution of the temperature field in time and space, and is the loss function directly embedded into the neural network as prior knowledge. The initial condition constraints mean that at the initial moment, i.e. t=0, the temperature distribution at each location inside the transformer must be equal to the known initial temperature field. The boundary condition constraints mean that at the transformer boundary, such as the winding surface and the tank wall, the temperature or heat flux must satisfy physical laws, such as the heat flux being proportional to the temperature difference at the heat dissipation boundary.

[0048] For example, inside a transformer, the temperature field T(t,x,y,z) satisfies the unsteady-state heat conduction equation with a heat source as follows: Where ρ is density, with units of kg / m³; c p is specific heat capacity, in J / (kg·K); k is thermal conductivity, in W / (m·K); q is internal heat source, such as copper loss or iron loss, in W / m³; ▽ refers to the Nabra operator, and ▽T represents the rate of change of temperature in all directions in space, that is, the direction and magnitude of heat flow.

[0049] Secondly, the physical parameters of the transformer, such as specific heat capacity, thermal conductivity, and thermal diffusivity, are encoded as learnable embedding vectors and introduced into the model as physical prior embedding vectors. This preserves the physical meaning of the physical parameters while allowing the model to fine-tune the prior parameters based on actual observation data, adapting to the actual parameter differences of different operating conditions and different transformer models. The physical parameters of the transformer are encoded as learnable embedding vectors and set as physical prior embedding vectors.

[0050] Finally, in the calculation of the physical constraint loss function, a residual-based attention mechanism is introduced. Higher weights are assigned to the physical equation constraint terms with larger residuals, and the contribution weights of different residual points to the total loss function are dynamically adjusted. The physical constraint loss value is finally calculated, which guides the neural network to prioritize fitting parts that do not conform to physical laws during the training process, thereby improving the constraint effect of physical constraints and obtaining a physical prior embedding vector that takes into account both data fitting and consistency with physical laws.

[0051] Furthermore, the physical constraint loss value = data fitting loss term + physical equation residual loss term + initial condition constraint loss term + boundary condition constraint loss term. Each loss term is balanced by configurable weights. For example, the weight of the data fitting loss term is fixed at 1, while the weights of the physical equation residual, initial condition, and boundary condition constraint loss terms can be dynamically adjusted according to the magnitude of each residual during training. This step embeds physical domain knowledge into the neural network encoding process, avoiding the erroneous results that easily occur in purely data-driven models that violate basic physical laws. At the same time, physical constraints reduce the model's dependence on large-scale labeled data and improve the model's generalization ability.

[0052] Based on the above step S20, an embedding vector incorporating prior knowledge of heat conduction physics can be obtained. Compared with the pure data-driven encoding method, the obtained physical prior embedding vector not only retains the information characteristics of the original monitoring data, but also satisfies the basic laws of transformer thermodynamics. It can provide a more physically realistic basic input for subsequent feature fusion and avoid unreasonable evaluation results that violate common sense physics.

[0053] S30: Based on the multi-source heterogeneous state data and the physical prior embedding vector, perform spatiotemporal graph Transformer fusion according to the transformer physical topology to obtain cross-modal fusion feature vector;

[0054] In this embodiment, based on the actual connection positions of the transformer's windings, core, tank, and other components, a topological graph structure is constructed with equipment components as nodes and physical heat conduction paths as edges. The monitoring features of each node in the multi-source heterogeneous state data are concatenated with the physical prior embedding vector to obtain the initial feature input of each node, which is then input into a graph perception Transformer. This spatiotemporal Transformer encoder uses a preset transformer physical topology to construct an attention graph. Through a space-time attention mechanism, it outputs a unified cross-modal fusion feature vector, achieving effective fusion of multi-source heterogeneous monitoring data and physical prior knowledge, while taking into account the spatiotemporal correlation and physical structural characteristics of the data.

[0055] Specifically, step S30 in the method includes:

[0056] Based on the actual physical structure of the transformer, a sparse and asymmetric dynamic graph matrix is ​​constructed. The actual physical structure of the transformer includes the connection relationship and energy transfer path between the core, high-voltage winding, low-voltage winding, oil passage and radiator.

[0057] The multi-source heterogeneous state data from different sensors are processed by independent encoding networks and mapped to a unified semantic space to obtain multimodal feature embeddings.

[0058] Based on the dynamic graph matrix and the multimodal feature embedding, cross-modal spatial-temporal attention computation of the graph-aware Transformer is performed to obtain the cross-modal fusion feature vector.

[0059] In this embodiment, firstly, based on the actual physical structure of the transformer, namely the connection relationship and energy transfer path between the core, high-voltage winding, low-voltage winding, oil passage and radiator, a topological relationship is constructed with each core component as a node and the heat conduction path as an edge, to obtain a sparse dynamic graph matrix that describes the physical connection of the transformer. This matrix limits attention calculation to be performed only between nodes with physical connections, avoiding invalid attention interference from nodes without physical connections.

[0060] Secondly, multi-source heterogeneous state data from different sources, including online monitoring, operation, environment, and ledgers, are input into independent modality coding branches. The original features of different dimensions and scales are mapped to a unified semantic space with consistent dimensions, resulting in multimodal feature embeddings for each modality. These multimodal feature embeddings are then concatenated with the physical prior embedding vectors obtained in step S20 to obtain the initial features for each topological node. For example, mapping the original features to a unified semantic space with consistent dimensions can be achieved using a multilayer perceptron, ensuring that different modal data can obtain encoding results adapted to their own feature distributions.

[0061] Finally, the initial features are input into the graph-aware Transformer encoder. The cross-modal spatial-temporal attention of the graph-aware Transformer refers to the simultaneous modeling of the spatial dimension of structural correlation and the capture of the temporal dimension of evolutionary laws in the embedding of multimodal features under the guidance of the dynamic graph matrix. The feature interaction weights between different sensors and at different times are calculated through the attention mechanism to obtain the cross-modal fusion feature vector.

[0062] Specifically, based on the dynamic graph matrix, attention weights are calculated only between physically connected nodes to mine temperature correlation features between different components; in the time attention calculation stage, the dynamic change features of monitoring data of the same component at different time steps are captured, and through multiple space-time attention calculations, a cross-modal fusion feature vector that integrates multi-source monitoring information and physical prior knowledge is finally obtained.

[0063] Based on the above step S30, the physical structure information of the transformer itself is fully utilized, and the spatiotemporal correlation of the monitoring data is effectively captured, realizing the accurate fusion of multi-source heterogeneous data and providing a high-quality feature foundation for subsequent multi-task health assessment.

[0064] S40: Based on the cross-modal fusion feature vector, perform multi-task health assessment and simultaneously obtain the health level classification probability distribution, hotspot temperature prediction value and remaining life prediction value;

[0065] In this embodiment, based on cross-modal fusion feature vectors, a health level classification branch, a hotspot temperature regression branch, and a remaining life regression branch are set. The health level classification branch maps the fused features to the probability distributions corresponding to four health levels: healthy, alert, abnormal, and faulty, and outputs the classification result of the current health level of the device. The hotspot temperature regression branch directly regresses and outputs the current value of the spatiotemporal hotspot temperature of the transformer based on the fused features and physical prior information. The remaining life regression branch outputs the prediction result of the remaining usable life of the transformer based on the aging features accumulated from the current features.

[0066] Specifically, step S40 in the method includes:

[0067] A multi-task learning architecture based on parameter sharing is adopted, which includes a health level classification branch, a hotspot temperature regression branch, and a remaining lifespan regression branch.

[0068] The cross-modal fusion feature vector is simultaneously input into the health level classification branch, the hotspot temperature regression branch, and the remaining lifespan regression branch;

[0069] The probability of each of the preset number of health levels is output through the health level classification branches and then merged into the health level classification probability distribution.

[0070] The current value of the spatiotemporal hot spot temperature of the transformer winding and its evolution trend within a preset time window are output through the hot spot temperature regression branch and set as the predicted hot spot temperature value.

[0071] The remaining service life of the output transformer is determined by the remaining lifetime regression branch and set as the predicted remaining lifetime value.

[0072] In this embodiment, a multi-task learning architecture based on parameter sharing is first adopted, which separates the output into three independent task branches. This approach leverages the commonalities between different tasks to improve overall evaluation performance and reduce the risk of overfitting that can easily occur with single-task training. Furthermore, task-specific branch networks adapt to the output characteristics of different tasks. The three branches correspond to the health level classification branch, the hotspot temperature regression branch, and the remaining lifespan regression branch, respectively.

[0073] Secondly, the cross-modal fusion feature vector obtained in step S30 is simultaneously input into the health level classification branch, the hotspot temperature regression branch, and the remaining lifespan regression branch to complete the inference calculation for different tasks.

[0074] Specifically, the health level classification branch maps the fused features to the probability outputs of a preset number of health levels through a fully connected layer, and merges them to obtain the final health level classification probability distribution. The current health level of the transformer can be directly determined based on the maximum probability, or the probability distribution can be output for maintenance personnel to refer to for decision-making.

[0075] For example, the health level classification branch is connected to two fully connected hidden layers and a Softmax output layer in sequence. Each fully connected hidden layer is followed by a ReLU activation function and batch normalization. Finally, the Softmax output layer outputs the probabilities corresponding to the four health levels: healthy, attention, abnormal, and fault, thus obtaining the health level classification probability distribution. The hidden layer dimension is adapted to the input fusion feature dimension to avoid overfitting due to excessive model parameters, while ensuring sufficient expressive power to fit the classification decision boundary.

[0076] For example, the training steps for the health level classification branch are as follows: First, the training dataset is divided into a training set and a validation set in an 8:2 ratio, and the classification loss is calculated using the cross-entropy loss function; after each training iteration, the classification accuracy is calculated on the validation set; when the validation set accuracy does not improve for 10 consecutive iterations, the training is terminated early to avoid overfitting; finally, the model parameters that perform best on the validation set are retained to complete the training of the classification branch.

[0077] The hot spot temperature regression branch outputs the current hot spot temperature at each location of the transformer winding through the regression output layer, as well as the hot spot temperature evolution trend within a preset time window. The above results are output as the hot spot temperature prediction value, which can help maintenance personnel to grasp the winding temperature change trend in advance and take timely heat dissipation or load reduction measures.

[0078] For example, the hotspot temperature regression branch also sequentially connects two fully connected hidden layers and a linear regression output layer. Each fully connected hidden layer is followed by a ReLU activation function and batch normalization, and finally outputs the hotspot temperature value corresponding to the prediction dimension. The output dimension is determined by the number of target prediction locations and the number of time steps. The training steps for the health level classification branch are as follows: calculate the regression loss using the mean squared error loss function, divide the training set and validation set in an 8:2 ratio, calculate the loss value on the validation set after each training iteration, and terminate the training early when the validation set loss does not decrease after 10 consecutive iterations, retain the model parameters with the minimum validation set loss, and complete the training of the hotspot temperature regression branch.

[0079] The remaining lifetime regression branch is based on the cumulative aging features extracted from the fusion features. It regresses the remaining available time of the transformer from its current state until it can no longer meet the requirements for safe operation. This is output as the remaining lifetime prediction value, providing data support for equipment rotation and maintenance planning.

[0080] For example, the remaining life regression branch also maps through a fully connected network to output a continuous scalar value, which is the predicted remaining service life of the transformer. This corresponding remaining life prediction result facilitates maintenance personnel in making advance plans for equipment replacement or maintenance. The remaining life regression branch also uses a fully connected network architecture to map the fused features into a single scalar output, corresponding to the remaining operational life of the transformer. During training, it also uses the mean squared error loss function and follows the same early stopping strategy as the hotspot temperature regression branch to ensure the model's generalization ability.

[0081] Based on this multi-task evaluation architecture, a multi-dimensional evaluation result of the transformer health status can be obtained synchronously with a single input, avoiding redundant computation caused by training and inference of multiple single-task models separately, while improving the evaluation accuracy of each task by utilizing the correlation information between tasks.

[0082] Furthermore, the substation equipment health status assessment method based on multi-source monitoring data fusion also includes:

[0083] Based on the cross-modal fusion feature vector, the component fault classification branch is executed to obtain the component-level fault type prediction results of the transformer, wherein the component-level fault types include core fault, winding fault, tap changer fault and insulation fault.

[0084] The component fault classification branch shares the cross-modal fusion feature vector with the health level classification branch, the hotspot temperature regression branch, and the remaining life regression branch, and is trained synchronously using a multi-task learning architecture.

[0085] In this embodiment, firstly, a component fault classification branch is added to the multi-task learning architecture. This branch also shares the cross-modal fusion features learned by the previous encoding and fusion modules, and only sets independent task-specific parameters in the output layer to adapt to the output requirements of component fault classification.

[0086] Secondly, the component fault classification branch, through a fully connected network and a Softmax output layer, outputs the probability distributions corresponding to different component-level fault types, such as core faults, winding faults, tap changer faults, and insulation faults, obtaining the final component-level fault type prediction results. This allows for further identification of the specific component and fault type after determining that the overall health status of the equipment is abnormal, providing clearer guidance for operation and maintenance and reducing the time cost of fault diagnosis. With the addition of this branch, all tasks are trained synchronously, and the supervision signals between different tasks can complement each other, further improving the expressive power of cross-modal fusion features and enhancing the overall performance of multi-task health assessment.

[0087] For example, the training component fault classification branch uses the cross-entropy loss function to calculate the classification loss. The training set and validation set are also divided in an 8:2 ratio. After each training iteration, the classification accuracy is calculated on the validation set. When the validation set accuracy does not improve after 10 consecutive iterations, an early stopping mechanism is triggered. The model parameters with the highest classification accuracy on the validation set are retained, and the training of the component fault classification branch is completed. This branch only outputs the prediction result when the device health level is judged as abnormal or faulty. If the overall health level is judged as healthy or alert, the result of this branch is not output, reducing unnecessary computational overhead.

[0088] Based on the multi-dimensional evaluation results output by step S40 above, the health status of the transformer can be comprehensively characterized from multiple levels, such as overall health level, key status parameters, remaining service life and specific fault types. This provides substation equipment operation and maintenance with sufficiently rich and accurate decision-making basis, avoiding the problem that outputting only a single evaluation result cannot meet the diverse needs of on-site operation and maintenance.

[0089] S50: Based on the model parameters in the multi-task health assessment, perform Bayesian uncertainty quantification to obtain the confidence intervals of the health level classification probability distribution, the hotspot temperature prediction value, and the remaining lifespan prediction value.

[0090] Bayesian uncertainty quantification refers to transforming the deterministic parameters of the original network into normally distributed random variables through variational inference. During the model inference phase, the parameters are sampled multiple times using Monte Carlo methods to obtain multiple prediction results. Based on the statistical characteristics of the prediction results, the confidence interval of each output result is calculated. This method does not require additional training data and only requires minor adjustments to the original model structure. It does not significantly increase the number of model parameters or inference time. This method can effectively quantify the uncertainty of model predictions, supplement confidence information for evaluation results, and further improve the safety of on-site operation and maintenance decisions.

[0091] In this embodiment of the application, Bayesian uncertainty quantization is performed based on the weights and other parameters of each layer of the neural network in the multi-task health assessment. By converting the deterministic network parameters into a probability distribution, multiple prediction results under different parameter sampling are obtained.

[0092] Based on the distribution statistics of the prediction results, the confidence intervals for the prediction results of health level classification, hotspot temperature prediction, and remaining life prediction can be calculated separately. The width of the confidence interval reflects the degree of uncertainty of the prediction results. If the confidence interval width is small, it means that the model has a high degree of confidence in the prediction results. If the confidence interval width is large, it means that the prediction results are highly uncertain. It is recommended that maintenance personnel further verify the results by combining other detection methods to avoid misjudgment caused by directly accepting uncertain prediction results.

[0093] Specifically, step S50 in the method includes:

[0094] The neural network weight parameters in the multi-task health assessment are set as random variables, and a Bayesian neural network is introduced.

[0095] Sensor noise and data acquisition error in the multi-source heterogeneous state data are modeled as random uncertainty, and the distribution variance of the random variables is modeled as cognitive uncertainty.

[0096] The superposition results of the cognitive uncertainty and the accidental uncertainty are calculated for the health level classification probability distribution, the hotspot temperature prediction value and the remaining life prediction value, respectively, to obtain their respective confidence intervals.

[0097] In this embodiment, firstly, the original deterministic neural network weight parameters in the multi-task health assessment, such as the weight matrix of the fully connected layer, are transformed into random variables that follow a Gaussian distribution. A Bayesian neural network is then constructed, and the standard neural network layer is replaced by the Bayesian neural network layer, so that the model output has a probability distribution, thereby quantifying cognitive uncertainty.

[0098] Secondly, the sources of uncertainty are modeled hierarchically. The measurement noise of the sensor itself and the data acquisition error caused by the interference of the field environment are modeled as random uncertainty. Random uncertainty stems from the inherent noise and measurement error of the data itself, such as sensor accuracy limitations and random environmental fluctuations, which cannot be eliminated by increasing the amount of data. The variance of the Bayesian parameter distribution is modeled as cognitive uncertainty. Cognitive uncertainty stems from the incompleteness of the model parameters and structure, such as insufficient training data or inappropriate selection of network depth, which can be reduced by adding more data or optimizing the model.

[0099] Finally, the total uncertainty is calculated for each output task. The variance of cognitive uncertainty and random uncertainty is superimposed to obtain the total variance. The confidence interval of the corresponding prediction result is calculated based on the total variance and the mean of the output prediction. The confidence level can be preset according to the on-site operation and maintenance requirements for decision security.

[0100] For example, for a 95% confidence level, the upper and lower limits of the confidence interval are the predicted mean plus or minus 1.96 times the total standard deviation, which can obtain a confidence interval that meets the preset confidence requirements. Different output results can have their confidence levels set independently, or the same confidence level can be set uniformly to adapt to the security requirements of different operation and maintenance scenarios.

[0101] Based on step S50 above, the confidence interval corresponding to each prediction result can be obtained. Maintenance personnel can judge the reliability of the prediction result based on the width of the confidence interval. For prediction results with high uncertainty, timely manual review can be arranged, effectively reducing the risk of misjudgment and improving the reliability and safety of substation equipment maintenance decisions. This step only transforms deterministic parameters into probability distributions on the existing network, without significantly modifying the model structure or adding additional labeled training data. It does not significantly increase the number of model parameters or inference time, and can supplement the evaluation results with confidence information at a relatively low cost, adapting to the implementation needs of field engineering applications.

[0102] S60: Based on the attention weight matrix in the spatiotemporal graph Transformer fusion and the contribution ratio of each equation term in the physical constraint loss value, perform physical attention attribution and generate an interpretability assessment report.

[0103] In this embodiment of the application, after completing the health status assessment and uncertainty quantification, an interpretable assessment report is further generated based on the intermediate results in the model calculation process, which solves the problem that the assessment results of traditional black-box AI models cannot be interpreted and are difficult to gain the trust of operation and maintenance personnel.

[0104] First, the attention weight matrix output by the spatial attention module in the graph-aware Transformer encoder is extracted. This weight directly reflects the contribution of different topological nodes, i.e., different device components, to the final evaluation result. The higher the weight, the greater the impact of the monitoring data of that component on the current evaluation result. Then, the contribution ratio of each physical equation term in the physical constraint loss function to the total loss is calculated. The higher the contribution ratio, the greater the deviation of the current device state from the physical law.

[0105] By combining the results of the two parts mentioned above, physical-guided attention attribution can be achieved: on the one hand, the key monitoring components and key monitoring periods that have the greatest impact on the current health status assessment results can be identified, and the source of the results can be clarified; on the other hand, by combining the contribution ratio of physical constraints, the physical laws that the assessment results conform to or deviate from can be explained, so that the assessment results have traceable physical evidence.

[0106] Finally, by integrating the results of multi-task evaluation, confidence intervals, and physical attention attribution, a structured and interpretable evaluation report is generated. This provides substation operation and maintenance personnel with a complete evaluation output that includes both accurate prediction results and clear interpretable evidence, assisting them in making quick and accurate decisions.

[0107] Specifically, step S60 in the method further includes:

[0108] The attention weight matrix is ​​extracted from the attention mechanism of each layer of the spatiotemporal graph Transformer fusion, the contribution score of each sensor channel in the multi-source heterogeneous state data to the evaluation decision is calculated, and a list of key sensor contributions is generated.

[0109] The physical constraint loss value is decomposed into the heat conduction term, convection term, and boundary condition term in the transformer heat conduction partial differential equation. The contribution ratio of each term to the physical constraint loss value is calculated to obtain the deviation location of the physical equation term.

[0110] Based on the list of contributions from the key sensors and the deviation location of the physical equation terms, an interpretability assessment report in natural language is generated. The interpretability assessment report also includes the current health level and its confidence interval, a description of the deviation from the expected state of the physical model, and operation and maintenance suggestions.

[0111] In this embodiment, firstly, attention weight matrices are extracted from the spatial attention and temporal attention modules of each layer of the spatiotemporal graph Transformer. The attention weights of the spatial and temporal dimensions are normalized and aggregated respectively. The contribution score of the monitoring data corresponding to each sensor acquisition channel to the final evaluation decision is calculated. The key sensor contribution list is generated according to the scores from high to low, presenting the monitoring data that has the greatest impact on the evaluation results, and helping maintenance personnel to quickly locate suspected abnormal components.

[0112] For example, the contribution score for the final evaluation decision can be calculated by weighting the spatial attention weights by the time dimension to obtain the total spatial attention score for each topological node. This total score is the spatial contribution score of the sensor data collected on the corresponding node. Then, the temporal attention weights are weighted by the spatial dimension to obtain the temporal contribution score of the monitoring data at each time step. Combining the temporal contribution distribution within the target evaluation period, the time interval of abnormal changes can be located. Finally, the spatial contribution score and the temporal contribution score are weighted and fused according to preset weights to obtain the final contribution score for each sensor channel. After the scores are sorted, a contribution list can be generated. The preset weights are adjusted according to the evaluation requirements. For example, when the requirement for abnormal time location is high, the weight of the temporal contribution score can be appropriately increased. When the requirement for abnormal component location is high, the weight of the spatial contribution score can be appropriately increased.

[0113] Secondly, the overall physical constraint loss value is decomposed according to the components of the heat conduction partial differential equation. The contribution ratio of the heat conduction term, convection term, and boundary condition term to the total physical constraint loss value is calculated separately. The higher the contribution ratio, the greater the deviation between the current actual state of the equipment and the law described by the physical term. This allows us to locate the deviation of the physical equation term and clarify which type of physical change characteristics the abnormal state conforms to.

[0114] For example, the contribution ratio can be calculated by dividing the absolute value of the loss component corresponding to each physical item by the sum of the absolute values ​​of the loss components of all physical items. The calculation result reflects the proportion of the deviation degree of each physical item to the total deviation degree, and can intuitively show the magnitude of the deviation contribution of each physical item without additional normalization processing.

[0115] Finally, the contribution list of key sensors and the deviation location of physical equation terms are integrated with the previously obtained multi-dimensional evaluation results and the confidence intervals corresponding to each result to generate an interpretable evaluation report in natural language. In addition to the core evaluation data, the report will also supplement the deviation description between the current equipment status and the expected status of the physical model, and provide corresponding operation and maintenance suggestions based on the evaluation results and deviations. This ensures that the entire evaluation process, from input data to output conclusions, has traceable and interpretable evidence, increases the trust of operation and maintenance personnel in the AI ​​evaluation results, and lowers the decision-making threshold.

[0116] Furthermore, the substation equipment health status assessment method based on multi-source monitoring data fusion also includes:

[0117] When a new batch of monitoring data arrives, incremental learning is performed to fine-tune and update the model parameters in the current multi-task health assessment based on the new batch of monitoring data.

[0118] The model in the multi-task health assessment deployed on the cloud server is used as the cloud teacher model;

[0119] Based on the cloud-based teacher model, the capabilities of the cloud-based teacher model are transferred to a lightweight neural network to be trained through knowledge distillation to obtain a lightweight student model, which is then deployed at the edge of the substation.

[0120] Real-time assessment is performed through cloud-edge collaboration between the cloud-based teacher model and the lightweight student model.

[0121] In this embodiment, the substation online monitoring system firstly collects and updates the equipment's operating data. As new data is constantly generated, the model fine-tunes its parameters on the new data through incremental learning. This allows the model to adapt to the slow changes in the equipment's operating status over time, without having to train from scratch with all historical data each time, thus significantly reducing the computational cost of model updates.

[0122] Secondly, cloud servers have sufficient computing power to support the inference and incremental update needs of complete large models, while the computing power of edge devices in substations is limited and cannot directly deploy large models. Therefore, knowledge distillation is used to transfer the knowledge learned by the complete teacher model in the cloud to a lightweight student model with fewer parameters, allowing the student model to approach the evaluation accuracy of the teacher model with lower computational overhead.

[0123] Finally, the lightweight student model trained is deployed to the edge of the substation. The edge can directly evaluate the locally collected monitoring data in real time. It only interacts with the cloud when the evaluation results are highly uncertain and need further verification, or when the model is updated. This satisfies the real-time requirements of on-site evaluation and can rely on the cloud's ability to continuously update the model, thus achieving efficient and stable operation of cloud-edge collaboration.

[0124] The cloud-based teacher model refers to a complete model deployed on a cloud server, containing a full multi-task evaluation architecture and all parameters. It leverages the ample computing power of the cloud to complete all processes, including incremental learning updates, Bayesian uncertainty quantification, and interpretability report generation. It also provides distilled soft labels for the lightweight student model, ensuring the student model can learn the generalization ability of the full large model. The lightweight student model, on the other hand, reduces the number of encoder layers in the spatiotemporal graph Transformer and compresses parameters by lowering the feature projection dimension, retaining only the core feature extraction and multi-task prediction structures. This satisfies the low inference latency requirements at the edge while preserving as much of the evaluation performance of the full large model as possible.

[0125] In summary, compared with existing technologies, this application achieves full interaction of multimodal features through the spatiotemporal graph Transformer fusion of multi-source heterogeneous monitoring data, and outputs multi-dimensional health assessment results simultaneously by combining multi-task joint learning, meeting diverse on-site operation and maintenance needs; it supplements confidence information for each assessment result through Bayesian uncertainty quantification, effectively reducing the risk of misjudgment; it generates interpretable assessment reports through physical attention attribution, solving the problem of the difficulty in implementing black-box models; and finally, it adapts to the long-term changes in the status of substation equipment through incremental learning and cloud-edge collaboration, meeting the real-time and updatable requirements for engineering implementation, comprehensively improving the accuracy, reliability, and practicality of substation equipment health status assessment.

[0126] In summary, the embodiments of this application have at least the following technical effects:

[0127] This application provides a substation equipment health status assessment method based on multi-source monitoring data fusion. First, the physical laws of transformer heat conduction are embedded into the neural network encoding process as physical constraint loss. A spatiotemporal graph Transformer is constructed based on the actual physical topology of the transformer to achieve cross-modal fusion of multi-source heterogeneous data. This utilizes prior physical knowledge to improve model generalization ability and ensures assessment accuracy through data-driven feature fusion. Second, a multi-task health assessment with shared parameters is adopted, simultaneously outputting health level classification probabilities, hotspot temperature predictions, and remaining life predictions, achieving multi-dimensional health status output. Third, a Bayesian neural network is used to model cognitive uncertainty and random uncertainty for model parameters and data noise, providing clear confidence intervals for each output result and reducing the risk of misjudgment. Finally, a natural language interpretable assessment report is generated through physical attention attribution, clarifying the contribution of each sensor data and physical equation term to the assessment results, meeting the power system's requirements for decision-making transparency. Simultaneously, incremental learning enables dynamic updates of model parameters, and knowledge distillation facilitates cloud-edge collaborative deployment, adapting to the real-time assessment needs of substation edge environments.

[0128] Through the above technical solutions, this application solves the problems of insufficient interpretability, single output dimension, lack of result confidence description, inability to dynamically update the model, and difficulty in adapting to edge deployment in existing assessment methods. It can provide transparent, comprehensive, and reliable health status assessment support for substation transformer operation and maintenance, effectively improve the intelligence level of substation equipment operation and maintenance, and ensure the stability and security of power grid operation.

[0129] Example 2, as Figure 3 As shown, based on the same inventive concept as the substation equipment health status assessment method using multi-source monitoring data fusion provided in Embodiment 1, this application also provides a substation equipment health status assessment system using multi-source monitoring data fusion, including:

[0130] Initial data acquisition module 11 is used to load multi-source heterogeneous state data of the transformer;

[0131] The neural network encoding module 12 is used to perform physical information neural network encoding based on the multi-source heterogeneous state data and the transformer heat conduction partial differential equation to obtain physical constraint loss value and physical prior embedding vector;

[0132] Feature vector fusion module 13 is used to perform spatiotemporal graph Transformer fusion based on the multi-source heterogeneous state data and the physical prior embedding vector, according to the transformer physical topology, to obtain cross-modal fusion feature vectors;

[0133] The multi-task health assessment module 14 is used to perform multi-task health assessment based on the cross-modal fusion feature vector, and simultaneously obtain the health level classification probability distribution, hotspot temperature prediction value and remaining life prediction value.

[0134] The Bayesian quantization execution module 15 is used to perform Bayesian uncertainty quantification based on the model parameters in the multi-task health assessment to obtain the confidence intervals of the health level classification probability distribution, the hotspot temperature prediction value, and the remaining life prediction value, respectively.

[0135] The evaluation report generation module 16 is used to perform physical attention attribution and generate an interpretability evaluation report based on the attention weight matrix in the spatiotemporal graph Transformer fusion and the contribution ratio of each equation term in the physical constraint loss value.

[0136] In one embodiment, the initial data acquisition module 11 is specifically used for:

[0137] Obtain online monitoring data, operational data, environmental data, and ledger data of the transformer;

[0138] The online monitoring data, the operational data, the environmental data, and the ledger data are denoised, missing value imputed, and standardized, and then merged into the multi-source heterogeneous state data.

[0139] In one embodiment, the neural network encoding module 12 is specifically used for:

[0140] The transformer heat conduction partial differential equation is used as the core of physical constraints, and initial condition constraints and boundary condition constraints are incorporated to construct a physical constraint loss function.

[0141] The physical parameters of the transformer are encoded into learnable embedding vectors, which are set as physical prior embedding vectors. The physical parameters include specific heat capacity, thermal conductivity and thermal diffusivity.

[0142] In the calculation of the physical constraint loss function, an attention mechanism based on residuals is introduced to dynamically adjust the contribution weight of different residual points to the total loss function, thereby obtaining the physical constraint loss value.

[0143] In one embodiment, the feature vector fusion module 13 is specifically used for:

[0144] Based on the actual physical structure of the transformer, a sparse and asymmetric dynamic graph matrix is ​​constructed. The actual physical structure of the transformer includes the connection relationship and energy transfer path between the core, high-voltage winding, low-voltage winding, oil passage and radiator.

[0145] The multi-source heterogeneous state data from different sensors are processed by independent encoding networks and mapped to a unified semantic space to obtain multimodal feature embeddings.

[0146] Based on the dynamic graph matrix and the multimodal feature embedding, cross-modal spatial-temporal attention computation of the graph-aware Transformer is performed to obtain the cross-modal fusion feature vector.

[0147] In one embodiment, the multi-task health assessment module 14 is specifically used for:

[0148] A multi-task learning architecture based on parameter sharing is adopted, which includes a health level classification branch, a hotspot temperature regression branch, and a remaining lifespan regression branch.

[0149] The cross-modal fusion feature vector is simultaneously input into the health level classification branch, the hotspot temperature regression branch, and the remaining lifespan regression branch;

[0150] The probability of each of the preset number of health levels is output through the health level classification branches and then merged into the health level classification probability distribution.

[0151] The current value of the spatiotemporal hot spot temperature of the transformer winding and its evolution trend within a preset time window are output through the hot spot temperature regression branch and set as the predicted hot spot temperature value.

[0152] The remaining service life of the output transformer is determined by the remaining lifetime regression branch and set as the predicted remaining lifetime value.

[0153] Furthermore, the substation equipment health status assessment method based on multi-source monitoring data fusion also includes:

[0154] Based on the cross-modal fusion feature vector, the component fault classification branch is executed to obtain the component-level fault type prediction results of the transformer, wherein the component-level fault types include core fault, winding fault, tap changer fault and insulation fault.

[0155] The component fault classification branch shares the cross-modal fusion feature vector with the health level classification branch, the hotspot temperature regression branch, and the remaining life regression branch, and is trained synchronously using a multi-task learning architecture.

[0156] Furthermore, in one embodiment of the application, the Bayesian quantization execution module 15 is specifically used for:

[0157] The neural network weight parameters in the multi-task health assessment are set as random variables, and a Bayesian neural network is introduced.

[0158] Sensor noise and data acquisition error in the multi-source heterogeneous state data are modeled as random uncertainty, and the distribution variance of the random variables is modeled as cognitive uncertainty.

[0159] The superposition results of the cognitive uncertainty and the accidental uncertainty are calculated for the health level classification probability distribution, the hotspot temperature prediction value and the remaining life prediction value, respectively, to obtain their respective confidence intervals.

[0160] Furthermore, in one embodiment, the evaluation report generation module 16 is specifically used for:

[0161] The attention weight matrix is ​​extracted from the attention mechanism of each layer of the spatiotemporal graph Transformer fusion, the contribution score of each sensor channel in the multi-source heterogeneous state data to the evaluation decision is calculated, and a list of key sensor contributions is generated.

[0162] The physical constraint loss value is decomposed into the heat conduction term, convection term, and boundary condition term in the transformer heat conduction partial differential equation. The contribution ratio of each term to the physical constraint loss value is calculated to obtain the deviation location of the physical equation term.

[0163] Based on the list of contributions from the key sensors and the deviation location of the physical equation terms, an interpretability assessment report in natural language is generated. The interpretability assessment report also includes the current health level and its confidence interval, a description of the deviation from the expected state of the physical model, and operation and maintenance suggestions.

[0164] Furthermore, the substation equipment health status assessment method based on multi-source monitoring data fusion also includes:

[0165] When a new batch of monitoring data arrives, incremental learning is performed to fine-tune and update the model parameters in the current multi-task health assessment based on the new batch of monitoring data.

[0166] The model in the multi-task health assessment deployed on the cloud server is used as the cloud teacher model;

[0167] Based on the cloud-based teacher model, the capabilities of the cloud-based teacher model are transferred to a lightweight neural network to be trained through knowledge distillation to obtain a lightweight student model, which is then deployed at the edge of the substation.

[0168] Real-time assessment is performed through cloud-edge collaboration between the cloud-based teacher model and the lightweight student model.

Claims

1. A method for assessing the health status of substation equipment based on multi-source monitoring data fusion, characterized in that, include: Load multi-source heterogeneous state data of the transformer; Based on the multi-source heterogeneous state data and the transformer heat conduction partial differential equation, physical information neural network encoding is performed to obtain physical constraint loss value and physical prior embedding vector; Based on the multi-source heterogeneous state data and the physical prior embedding vector, spatiotemporal graph Transformer fusion is performed according to the physical topology of the transformer to obtain cross-modal fusion feature vectors; Based on the cross-modal fusion feature vector, a multi-task health assessment is performed to simultaneously obtain the health level classification probability distribution, hotspot temperature prediction value, and remaining life prediction value. Based on the model parameters in the multi-task health assessment, Bayesian uncertainty quantification is performed to obtain the confidence intervals of the health level classification probability distribution, the hotspot temperature prediction value, and the remaining lifespan prediction value. Based on the attention weight matrix in the spatiotemporal graph Transformer fusion and the contribution ratio of each equation term in the physical constraint loss value, physical attention attribution is performed to generate an interpretability assessment report.

2. The substation equipment health status assessment method based on multi-source monitoring data fusion as described in claim 1, characterized in that, The multi-source heterogeneous state data of the loaded transformer includes: Obtain online monitoring data, operational data, environmental data, and ledger data of the transformer; The online monitoring data, the operational data, the environmental data, and the ledger data are denoised, missing value imputed, and standardized, and then merged into the multi-source heterogeneous state data.

3. The substation equipment health status assessment method based on multi-source monitoring data fusion as described in claim 1, characterized in that, Based on the aforementioned multi-source heterogeneous state data and the transformer heat conduction partial differential equation, physical information neural network encoding is performed to obtain physical constraint loss values ​​and physical prior embedding vectors, including: The transformer heat conduction partial differential equation is used as the core of physical constraints, and initial condition constraints and boundary condition constraints are incorporated to construct a physical constraint loss function. The physical parameters of the transformer are encoded into learnable embedding vectors, which are set as physical prior embedding vectors. The physical parameters include specific heat capacity, thermal conductivity and thermal diffusivity. In the calculation of the physical constraint loss function, an attention mechanism based on residuals is introduced to dynamically adjust the contribution weight of different residual points to the total loss function, thereby obtaining the physical constraint loss value.

4. The substation equipment health status assessment method based on multi-source monitoring data fusion as described in claim 1, characterized in that, Based on the multi-source heterogeneous state data and the physical prior embedding vector, spatiotemporal graph Transformer fusion is performed according to the transformer physical topology to obtain cross-modal fusion feature vectors, including: Based on the actual physical structure of the transformer, a sparse and asymmetric dynamic graph matrix is ​​constructed. The actual physical structure of the transformer includes the connection relationship and energy transfer path between the core, high-voltage winding, low-voltage winding, oil passage and radiator. The multi-source heterogeneous state data from different sensors are processed by independent encoding networks and mapped to a unified semantic space to obtain multimodal feature embeddings. Based on the dynamic graph matrix and the multimodal feature embedding, cross-modal spatial-temporal attention computation of the graph-aware Transformer is performed to obtain the cross-modal fusion feature vector.

5. The substation equipment health status assessment method based on multi-source monitoring data fusion as described in claim 1, characterized in that, Based on the cross-modal fusion feature vector, a multi-task health assessment is performed, simultaneously obtaining the health level classification probability distribution, hotspot temperature prediction, and remaining life expectancy prediction, including: A multi-task learning architecture based on parameter sharing is adopted, which includes a health level classification branch, a hotspot temperature regression branch, and a remaining lifespan regression branch. The cross-modal fusion feature vector is simultaneously input into the health level classification branch, the hotspot temperature regression branch, and the remaining lifespan regression branch; The probability of each of the preset number of health levels is output through the health level classification branches and then merged into the health level classification probability distribution. The current value of the spatiotemporal hot spot temperature of the transformer winding and its evolution trend within a preset time window are output through the hot spot temperature regression branch and set as the predicted hot spot temperature value. The remaining service life of the output transformer is determined by the remaining lifetime regression branch and set as the predicted remaining lifetime value.

6. The substation equipment health status assessment method based on multi-source monitoring data fusion as described in claim 1, characterized in that, Also includes: Based on the cross-modal fusion feature vector, the component fault classification branch is executed to obtain the component-level fault type prediction results of the transformer, wherein the component-level fault types include core fault, winding fault, tap changer fault and insulation fault. The component fault classification branch shares the cross-modal fusion feature vector with the health level classification branch, the hotspot temperature regression branch, and the remaining life regression branch, and is trained synchronously using a multi-task learning architecture.

7. The substation equipment health status assessment method based on multi-source monitoring data fusion as described in claim 1, characterized in that, Based on the model parameters in the multi-task health assessment, Bayesian uncertainty quantification is performed to obtain the confidence intervals for the health level classification probability distribution, the predicted hotspot temperature, and the predicted remaining lifespan, including: The neural network weight parameters in the multi-task health assessment are set as random variables, and a Bayesian neural network is introduced. Sensor noise and data acquisition error in the multi-source heterogeneous state data are modeled as random uncertainty, and the distribution variance of the random variables is modeled as cognitive uncertainty. The superposition results of the cognitive uncertainty and the accidental uncertainty are calculated for the health level classification probability distribution, the hotspot temperature prediction value and the remaining life prediction value, respectively, to obtain their respective confidence intervals.

8. The substation equipment health status assessment method based on multi-source monitoring data fusion as described in claim 1, characterized in that, Based on the attention weight matrix in the spatiotemporal graph Transformer fusion and the contribution ratio of each equation term in the physical constraint loss value, physical attention attribution is performed to generate an interpretability assessment report, including: The attention weight matrix is ​​extracted from the attention mechanism of each layer of the spatiotemporal graph Transformer fusion, the contribution score of each sensor channel in the multi-source heterogeneous state data to the evaluation decision is calculated, and a list of key sensor contributions is generated. The physical constraint loss value is decomposed into the heat conduction term, convection term, and boundary condition term in the transformer heat conduction partial differential equation. The contribution ratio of each term to the physical constraint loss value is calculated to obtain the deviation location of the physical equation term. Based on the list of contributions from the key sensors and the deviation location of the physical equation terms, an interpretability assessment report in natural language is generated. The interpretability assessment report also includes the current health level and its confidence interval, a description of the deviation from the expected state of the physical model, and operation and maintenance suggestions.

9. The substation equipment health status assessment method based on multi-source monitoring data fusion as described in claim 1, characterized in that, Also includes: When a new batch of monitoring data arrives, incremental learning is performed to fine-tune and update the model parameters in the current multi-task health assessment based on the new batch of monitoring data. The model in the multi-task health assessment deployed on the cloud server is used as the cloud teacher model; Based on the cloud-based teacher model, the capabilities of the cloud-based teacher model are transferred to a lightweight neural network to be trained through knowledge distillation to obtain a lightweight student model, which is then deployed at the edge of the substation. Real-time assessment is performed through cloud-edge collaboration between the cloud-based teacher model and the lightweight student model.

10. A substation equipment health status assessment system based on multi-source monitoring data fusion, characterized in that, The substation equipment health status assessment method for performing multi-source monitoring data fusion as described in any one of claims 1-9 includes: The initial data acquisition module is used to load multi-source heterogeneous state data of the transformer; The neural network encoding module is used to perform physical information neural network encoding based on the multi-source heterogeneous state data and the transformer heat conduction partial differential equation to obtain physical constraint loss value and physical prior embedding vector. The feature vector fusion module is used to perform spatiotemporal graph Transformer fusion based on the multi-source heterogeneous state data and the physical prior embedding vector, according to the transformer physical topology, to obtain cross-modal fused feature vectors. The multi-task health assessment module is used to perform multi-task health assessment based on the cross-modal fusion feature vector, and simultaneously obtain the health level classification probability distribution, hotspot temperature prediction value and remaining life prediction value. The Bayesian quantization execution module is used to perform Bayesian uncertainty quantification based on the model parameters in the multi-task health assessment, and to obtain the confidence intervals of the health level classification probability distribution, the hotspot temperature prediction value, and the remaining life prediction value, respectively. The evaluation report generation module is used to perform physical attention attribution and generate an interpretable evaluation report based on the attention weight matrix in the spatiotemporal graph Transformer fusion and the contribution ratio of each equation term in the physical constraint loss value.

Images